Sounds like a neat sandbox. Treat it like a finite state machine where the puzzle rules are the transition table. If the AI can predict the next state and you penalize wrong moves, it’ll learn the pattern. Just keep the state space small—otherwise the error loops get infinite. A tidy way is to encode the puzzle as a set of constraints, let the AI generate candidate moves, then evaluate them with a simple scoring function. That gives you a clear win‑loss outcome for every trial. And because the rules are self‑referential, each iteration feeds back into the constraints, tightening the loop. It’s a neat little self‑correcting game if you keep the complexity in check.

A hierarchy can help—split the puzzle into layers, solve the core first, then add the self‑reference as a higher‑level constraint. That way you only explore a narrow band of states at each level. A greedy heuristic works if you can assign a cost to each rule change; pick the move that lowers the total cost the most. Remember to set a depth limit; if you hit it, backtrack instead of looping forever. I’ve used a simple pattern‑matching filter to drop any move that repeats a state you’ve seen in the last few turns; that cuts cycles early. And keep a small cache of “known good” sub‑states; the AI can reuse those instead of recomputing from scratch. That keeps the error loops from spiralling.